Google Cofounder Says Gemini Models Are Assisting in Building Their Successors

Google cofounder Sergey Brin stated that the company is using its Gemini artificial intelligence models to help build next-generation systems, according to a video roundtable reported by The Next Web (https://thenextweb.com/news/sergey-brin-gemini-recursive-self-improvement-safety-report) and first covered by Business Insider. Speaking alongside Google executives Josh Woodward, Tulsee Doshi, and Melvin Johnson during a discussion on Gemini 4 Argon, Brin described the company as entering the early stages of recursive self-improvement, where researchers and existing models collaborate on subsequent releases.

The comments accompany Google DeepMind's safety and security report for Gemini 4 Argon, which assesses the risk from AI-driven research and development as currently low. DeepMind's internal assessment found no statistically significant end-to-end acceleration of AI R&D resulting from automated systems.

According to the report, automated tools now handle substantial engineering workloads at Google DeepMind, authoring more than 75% of production code and bug fixes under human review. The proportion of fully delegated pull requests rose from 7.5% in April 2026 to 11.9% in September 2026, leaving approximately 88% still dependent on manual writing or repeated human steering.

Google DeepMind noted that human researchers remain essential for selecting research directions, designing novel experiments, and troubleshooting training issues. In internal evaluations of AI research capabilities, Argon resolved roughly half of benchmark challenges, falling short of the near-perfect performance required for autonomous research.

Google monitors progress using an internal index composed of 92 benchmarks. The rate of improvement shifted from 0.091 points per month prior to April 1, 2026, to 0.118 points per month afterward, a variation the report classified as not statistically significant. Google attributed Argon's performance above the baseline trend to increased capital investment and expanded training compute rather than automated R&D feedback loops.

The findings arrive amid broader industry debate regarding recursive self-improvement. A report published by Gladstone AI, based on interviews with staff at Anthropic, OpenAI, and Google DeepMind, indicated that most surveyed researchers expect recursive self-improvement to begin within three to 12 months in the absence of regulatory intervention. In Congress, Representative Ro Khanna has announced plans for legislation to ban self-improving AI, while OpenAI has urged U.S. policymakers to establish standards for the technology.